Multi-Branch Visual Fingerprinting & Probabilistic Matching System
A production-ready, multi-branch object identity system deployed on Google Cloud Platform. The system ingests real-world images and reconstructs persistent object identity using complementary visual signals, uncertainty-aware evidence fusion, and explainable AI.
This is not image search. This system reasons about physical object continuity over time.
- Robust to angle, lighting, damage, occlusion, and partial views
- Tracks object identity evolution across sightings
- Probabilistic, uncertainty-aware matching
- Fully explainable results (heatmaps + natural language)
- Designed for Cloud Run + Vertex AI production deployment
“We do not search images. We reconstruct physical object identity using multi-signal intelligence.”
From ingestion → identity reconstruction → ranking → explainability
- ViT patch variance (latent manufacturing noise)
- CLIP image embeddings
- Gemini Vision understanding
- MediaPipe geometry
- Custom ghost signals
- Mask generation
- Edge & depth priors
- Imagen inpainting
- Completion embeddings
- Void signatures (128D)
- Structural absence reasoning
- Vertex AI Multimodal Embeddings
Runtime: FastAPI on Cloud Run
Storage: GCS, Firestore
AI/ML: Vertex AI (Gemini, Multimodal Embeddings, Imagen), MediaPipe, PyTorch, TFP
GET /health
POST /analyze
POST /feedback
Enable:
- Cloud Run
- Cloud Storage
- Firestore
- Vertex AI
Use Cloud Run with environment variables for GCS, Gemini, Imagen, and embeddings.
objects/{object_id}
sightings/{sighting_id}
fusion/reliability
- CPU Cloud Run works; GPU optional
- 2–4Gi memory recommended
- Least-privilege IAM
- Signed URLs if private
MIT or Apache-2.0